Overall Objectives of

Model Predictive Control

1. Prevent violations of input and output constraints.

Chapter 20

2. Drive some output variables to their optimal set

points, while maintaining other outputs within
specified ranges.
3. Prevent excessive movement of the input variables.
4. If a sensor or actuator is not available, control as
much of the process as possible.

1
 Model Predictive Control: Basic Concepts

1. Future values of output variables are predicted using a dynamic

model of the process and current measurements.
• Unlike time delay compensation methods, the predictions are
Chapter 20

made for more than one time delay ahead.

2. The control calculations are based on both future predictions and
current measurements.
3. The manipulated variables, u(k), at the k-th sampling instant are
calculated so that they minimize an objective function, J.
• Example: Minimize the sum of the squares of the deviations
between predicted future outputs and specific reference trajectory.
• The reference trajectory is based on set points calculated using
RTO.
4. Inequality & equality constraints, and measured disturbances are
included in the control calculations.
5. The calculated manipulated variables are implemented as set
point for lower level control loops. (cf. cascade control).
2
 Model Predictive Control: Calculations

1. At the k-th sampling instant, the values of the manipulated

variables, u, at the next M sampling instants, {u(k), u(k+1), …,
u(k+M -1)} are calculated.
• This set of M “control moves” is calculated so as to minimize
Chapter 20

the predicted deviations from the reference trajectory over the

next P sampling instants while satisfying the constraints.
• Typically, an LP or QP problem is solved at each sampling
instant.
• Terminology: M = control horizon, P = prediction horizon
2. Then the first “control move”, u(k), is implemented.
3. At the next sampling instant, k+1, the M-step control policy is
re-calculated for the next M sampling instants, k+1 to k+M, and
implement the first control move, u(k+1).
4. Then Steps 1 and 2 are repeated for subsequent sampling
instants.
Note: This approach is an example of a receding horizon
approach.
3
Chapter 20

Figure 20.2 Basic concept for Model Predictive Control

4
 When Should Predictive Control be Used?

1. Processes are difficult to control with standard PID

algorithm (e.g., large time constants, substantial time
delays, inverse response, etc.
Chapter 20

2. There is significant process interactions between u

and y.
• i.e., more than one manipulated variable has a significant effect on
an important process variable.

3. Constraints (limits) on process variables and

manipulated variables are important for normal control.
Terminology:
• y ↔ CV, u ↔ MV, d ↔ DV

5
 Model Predictive Control Originated in 1980s

• Techniques developed by industry:

1. Dynamic Matrix Control (DMC)
Chapter 20

• Shell Development Co.: Cutler and Ramaker (1980),

• Cutler later formed DMC, Inc.
• DMC acquired by Aspentech in 1997.

2. Model Algorithmic Control (MAC)

• ADERSA/GERBIOS, Richalet et al. (1978) in France.
• Over 5000 applications of MPC since 1980
Reference: Qin and Badgwell, 1998 and 2003).

6
 Dynamic Models for
Model Predictive Control
• Could be either:
1. Physical or empirical (but usually empirical)
Chapter 20

2. Linear or nonlinear (but usually linear)

• Typical linear models used in MPC:

1. Step response models
2. Transfer function models
3. State-space models

• Note: Can convert one type of linear model

(above) to the other types.

7
 Discrete Step Response Models

Consider a single input, single output process:

Chapter 20

u Process y

where u and y are deviation variables (i.e., deviations

from nominal steady-state values).

8
 Prediction for SISO Models:
Example: Step response model

N 1
y(k +1 )  y0   Si u( k  i  1 )  S N u( k  N  1 ) (20-1)
i 1
Chapter 20

Si = the i-th step response coefficient

N = an integer (the model horizon)
y0 = initial value at k=0

Figure 7.14. Unit Step

Response

9
 Prediction for SISO Models:
Example: Step response model
N 1
y(k +1 )  y0   Si u( k  i  1 )  S N u( k  N  1 ) (20-1)
Chapter 20

i 1

• If y0=0, this one-step-ahead prediction can be obtained from

Eq. (20-1) by replacing y(k+1) with yˆ ( k  1)
N 1
yˆ (k  1)   Si u(k  i  1)  S N u (k  N  1) (20  6)
i 1

• Equation (20-6) can be expanded as:

N 1
yˆ (k  1)  S1u (k )   Si u(k  i  1)  S N u (k  N  1)
i 2
Effect of current
control action Effect of past control actions

10
 Prediction for SISO Models:
(continued)
Similarly, the j-th step ahead prediction is Eq. 20-10:
j N 1
yˆ (k  j )   Si u(k  j  i)   Si u(k  j  i)  S N u (k  j  N )
Chapter 20

i 1 i  j 1
Effects of current and Effects of past
future control actions control actions

Define the predicted unforced response as:

N 1
yˆ (k  j )
o
 Si u(k  j  i)  S N u (k  j  N ) (20  11)
i= j+1

and can write Eq. (20-10) as:

j
yˆ (k  j )   i
S u ( k  j  i )  ˆ
y o
(k  j ) (20  12)
i 1 11
 Vector Notation for Predictions

Define vectors:

Yˆ (k + 1 ) col [ ˆy(k + 1 ), ˆy(k + 2 ), , ˆy(k + P)] (20-16)

Chapter 20

Yˆ o (k + 1 ) col [ ˆy o (k + 1 ), ˆy o (k + 2 ), , ˆy o (k + P)] (20-17)

ΔU(k) col [ Δu(k),Δu(k +1 ), ,Δu(k + M - 1 )] (20-18)

The model predictions in Eq. (20-12) can be written as:

Yˆ (k + 1 ) = SΔU (k)  Yˆ o (k + 1 ) (20-19)

12
 Dynamic Matrix Model

The model predictions in Eq. (20-12) can be

written as:
Yˆ (k + 1 ) = SΔU (k)  Yˆ o (k + 1 ) (20-19)
Chapter 20

where S is the P x M dynamic matrix:

 S1 0 0 
 S2 S1 0 
 
 0 
 
S  SM S M -1 S1  (20-20)
 S M +1 SM S2 
 
 
 
 SP S P-1 S P-M +1 
13
 Bias Correction

• The model predictions can be corrected by utilizing the

latest measurement, y(k).
• The corrected prediction is defined to be:
Chapter 20

y(k + j) ˆy(k + j)+ [y(k) - ˆy(k)] (20-23)

• Similarly, adding this bias correction to each prediction in

(20-19) gives:
Y (k + 1 ) = S U (k)  Yˆ o (k + 1 )+ [y(k) - ˆy(k)] 1 (20-24)
where Y (k +1 ) is defined as:

Y (k +1 ) col [y(k +1 ), y(k + 2 ), , y(k + P)] (20-25)

14
 EXAMPLE 20.4
The benefits of using corrected predictions will be illustrated by a simple
example, the first-order plus-time-delay model of Example 20.1:

Y(s) 5e-2s
= (20-26)
U(s) 15s  1
Assume that the disturbance transfer function is identical to the
Chapter 20

process transfer function, Gd(s)=Gp(s). A unit step change in u

occurs at time t=2 min and a step disturbance, d=0.15, occurs at
t=8 min. The sampling period is t= 1 min.
(a) Compare the process response y(k) with the predictions that
were made 15 steps earlier based on a step response model with
N=80. Consider both the corrected prediction
(b) Repeat part (a) for the situation where the step response coefficients
are calculated using an incorrect model:

Y(s) 4e-2 s
= (20-27)
U(s) 20s  1

15
Chapter 20

Figure 20.6 Without model error. 16

Chapter 20

Figure 20.7 With model error.

17
Chapter 20

Figure 20.10 Input blocking.

18
 Reference Trajectory for MPC
Reference Trajectory
• A reference trajectory can be used to make a gradual transition
to the desired set point.
• The reference trajectory Yr can be specified in several different
Chapter 20

ways. Let the reference trajectory over the prediction horizon P

be denoted as:

Yr (k + 1 ) col[yr (k + 1 ), yr (k + 2 ), , yr (k + P)] (20-47)

where Yr is an mP vector where m is the number of outputs.

Exponential Trajectory from y(k) to ysp(k)

A reasonable approach for the i-th output is to use:
yi,r (k+j) = (ai) j yi (k) + [1 - (ai) j] yi,sp (k) (20-48)

for i=1,2,…, m and j=1, 2, …, P.

19
 MPC Control Calculations

• The control calculations are based on minimizing the predicted

deviations between the reference trajectory.
Chapter 20

• The predicted error is defined as:

ˆ (k + 1 )
E Yr (k + 1 )  Y (k + 1 ) (20-50)

where Y (k + 1 ) is the corrected prediction defined in (20-37).

Similarly, the predicted unforced error , Eˆ o (k + 1 ), is defined as:

ˆ o (k + 1 )
E Yr (k + 1 )  Y o (k + 1 ) (20-51)
• Note that all of the above vectors are of dimension, mP.
• The objective of the control calculations is to calculate the control
policy for the next M time intervals:

ΔU (k) col [Δu(k),Δu(k +1 ), ,Δu(k + M - 1 )] (20-18)

20
 MPC Performance Index

• The rM-dimensional vector U(k) is calculated so as to minimize:

a. The predicted errors over the prediction horizon, P.
Chapter 20

b. The size of the control move over the control horizon, M.

• Example: Consider a quadratic performance index:
ˆ ( k  1 )T Q E
min J  E ˆ ( k  1 )  U ( k )T R U ( k ) (20 - 54)
U ( k )

where Q is a positive-definite weighting matrix and R is a

positive semi-definite matrix.

Both Q and R are usually diagonal matrices with positive

diagonal elements.

The weighting matrices are used to weight the most important

outputs and inputs (cf. Section 20.6).
21
 MPC Control Law: Unconstrained Case
•The MPC control law that minimizes the objective function in
Eq. (20-54) can be calculated analytically,

U (k) = ( S T Q S  R ) -1 S T Q Eˆ o (k +1 ) (20-55)

Chapter 20

where S is the dynamic matrix defined in (20-41).

• This control law can be written in a more compact form,
 mP
ˆ o (k +1 )
U(k)= Kc E (20-56)
where controller gain matrix Kc is defined to
be:
Kc ( ST Q S  R ) -1 S T Q (20-57)
• Note that Kc can be evaluated off-line, rather than on-line, provided
that the dynamic matrix S and weighting matrices, Q and R, are
constant.

• The calculation of Kc requires the inversion of an rM x rM matrix

where r is the number of input variables and M is the control horizon. 22
 MPC Control Law:
Receding Horizon Approach
• MPC control law:
ˆ o (k +1 )
U(k)= Kc E (20-56)
Chapter 20

where:
ΔU (k) col [Δu(k),Δu(k +1 ), ,Δu(k + M - 1 )] (20-18)
r  mP
• Note that the controller gain matrix, Kc, is an rM x mP matrix.
.

• In the receding horizon control approach, only the first step of the
M-step control policy, u(k), in (20-18) is implemented.
ˆ o (k +1 )
u(k)= Kc1 E (20-58)
where matrix Kc1 is defined to be the first r rows of Kc.
Thus, Kc1 has dimensions of r x mP.

23
 Selection of Design Parameters
Model predictive control techniques include a number of
design parameters:

N: model horizon
Chapter 20

t: sampling period

P: prediction horizon (number of predictions)
M: control horizon (number of control moves)
Q: weighting matrix for predicted errors (Q > 0)
R: weighting matrix for control moves (R > 0)

24
 Selection of Design Parameters (continued)
1. N and t
These parameters should be selected so that N t > open-loop
settling time. Typical values of N:
30 < N < 120
Chapter 20

2. Prediction Horizon, P
Increasing P results in less aggressive control action
Set P = N + M
3. Control Horizon, M
Increasing M makes the controller more aggressive and increases
computational effort, typically:
5 < M < 20
4. Weighting matrices Q and R
Diagonal matrices with largest elements corresponding to most
important variables

25
 Example 20.5: set-point responses

e-s
G(s) =
(10 s +1)(5s +1)
Chapter 20

26
 Example 20.5: disturbance responses
Chapter 20

27